Dynamic Defense Policies Learning Through Soft Actor-Critic for Soccer Robot Goalkeepers

This paper implements a dynamic goalkeeper defense policy based on Deep Reinforcement Learning (DRL) for a soccer robot. This study proposed a multi-directional training approach to enable soccer goalkeepers to learn defense strategies for acting an appropriate manner. This paper using the Soft Actor-Critic (SAC) as DRL architecture and the Gazebo simulator as an environment simulation platform. Utilizing progressive learning with good reward function design to enable soccer robot goalkeepers to explore and learn in a simulated environment. It efficiently completed the better defense task. Finally, the correctness of the defense policy is evaluated by testing and validating it in a multi-directional area. The experimental results demonstrated the designed SAC containing the superior performance for successful defense rates in this paper.